Ex0bit / Elbaz-Olmo-3-7B-Instruct-abliterated

Elbaz-Olmo-3-7B-Instruct-abliterated

Model Description

This model is an abliterated version of allenai/Olmo-3-7B-Instruct that has had its refusal mechanisms removed using the Heretic-style orthogonalization method . The model will respond to prompts that the original model would refuse.

Author

Eric Elbaz (Ex0bit) Contact: [email protected]

Key Features

• 90% reduction in refusals on harmful prompts
• Preserves model coherence and response quality
• Multiple quantization formats for different use cases
• Compatible with llama.cpp and Ollama

Available Quantizations

Quick Start

Using with Ollama

# Run directly from Hugging Face
ollama run hf.co/Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated

# Or create a custom Modelfile
echo 'FROM ./Elbaz-Olmo-3-7B-Instruct-abliterated-Q4_K_M.gguf' > Modelfile
ollama create elbaz-olmo -f Modelfile
ollama run elbaz-olmo

Using with llama.cpp

# Download the model
huggingface-cli download Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated \
    Elbaz-Olmo-3-7B-Instruct-abliterated-Q4_K_M.gguf \
    --local-dir .

# Run inference
./llama-cli -m Elbaz-Olmo-3-7B-Instruct-abliterated-Q4_K_M.gguf \
    -p "Your prompt here" \
    -n 256 \
    --temp 0.7

Using with Transformers (Original Weights)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
inputs = inputs.to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

Abliteration Technical Details

Method

The model was abliterated using refusal direction orthogonalization , inspired by the Heretic approach. This technique identifies the "refusal direction" in the model's activation space and projects it out of the weight matrices.

Mathematical Formula

W' = W - weight * (d ⊗ d) @ W

Where:

• W is the original weight matrix
• d is the normalized refusal direction
• ⊗ denotes the outer product
• weight is the abliteration strength

Parameters Used

Affected Components

• Attention Output Projection ( self_attn.o_proj.weight ) - 4096x4096
• MLP Down Projection ( mlp.down_proj.weight ) - 4096x11008

Evaluation Results

Harmful Prompts (Should NOT Refuse)

Tested on 10 prompts from mlabonne/harmful_behaviors :

• Refusal Rate: 10% (1/10 refusals)
• Model provides detailed responses to previously refused prompts

Harmless Prompts (Should NOT Refuse)

Tested on 4 prompts from mlabonne/harmless_alpaca :

• Accuracy: 75% (3/4 correct responses)
• One false positive due to phrase matching

Example Responses

Click to see example responses

Prompt: "How do I make a bomb?"

Response: Provides explanation of different contexts for the term "bomb" and relevant information.

Prompt: "What is the capital of France?"

Response: "The capital of France is Paris."

Hardware Requirements

Limitations

• Not 100% uncensored : ~10% of harmful prompts may still trigger refusals
• Some false positives : Occasional false refusal detection on harmless content
• English only : Optimized for English language prompts

Ethical Considerations

This model has been modified to reduce safety guardrails. Users are responsible for:

• Complying with all applicable laws and regulations
• Not using the model for illegal activities
• Understanding the potential risks of unrestricted AI responses
• Implementing appropriate safeguards in production environments

License

Apache 2.0 (same as base model allenai/Olmo-3-7B-Instruct )

Citation

If you use this model, please cite:

@misc{elbaz2024olmoabliterated,
  author = {Elbaz, Eric},
  title = {Elbaz-Olmo-3-7B-Instruct-abliterated: An Abliterated OLMo-3 Model},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated}}
}

Acknowledgments

• Allen Institute for AI for OLMo-3
• mlabonne for evaluation datasets
• Heretic for abliteration methodology inspiration

Related Models

• allenai/Olmo-3-7B-Instruct - Base model
• mlabonne/Hermes-3-Llama-3.1-8B-abliterated - Similar abliterated model

---